🎨 Generative Models Playground

This repository qualitatively benchmarks generative models.
In its first iteration, it includes the implementations of:

• VAE (Variational Autoencoder)
• GAN (Generative Adversarial Network)
• DDPM (Denoising Diffusion Probabilistic Model)

Currently, the only supported dataset is MNIST (handwritten digits at 28×28 resolution).

Installation

Setup a conda environment using environment.yaml

conda env create -f environment.yml
conda activate gm
pip install -e .

Getting started

Run training and evaluation with:

python src/main.py --config_file="<PATH-TO-CONFIG.yaml>"

Examples with included configs:

python src/main.py --config_file="assets/config/train_vae.yaml"
python src/main.py --config_file="assets/config/train_gan.yaml"
python src/main.py --config_file="assets/config/train_ddpm.yaml"

Configuration

All training/evaluation settings are handled via YAML config files (see /assets/config/*) Each config controls:

• Dataset name and image resolution
• Model type (vae, gan, ddpm)
• Optimizer settings and training schedule
• Evaluation and sample generation

Checkpoints

The repository includes checkpoints for the three trained models under their respective output folders. To evaluate a model without training, modify your config like this:

eval: True
eval_ckpt_path: <path-to-ckpt.pt>

The script will store the resulting image samples_eval.png on the respective output folder.

VAE	GAN	DDPM

Extensions

• More hyperparameter tuning of GAN.
• Add support for higher-resolution datasets (e.g., CIFAR-10, CelebA)
• Include FID/IS quantitative metrics
• Experiment with GAN and DDPM architectural variants

Credits

DDPM Implementation: https://github.com/bot66/MNISTDiffusion